Intelligent optimization device for AI auxiliary radiotherapy dose

Abstract

The invention discloses an AI auxiliary radiotherapy dose intelligent optimization device, which comprises a multi-modal image fusion module, a dose distribution prediction module, a dynamic adaptive optimization module and a prognosis model integration module, and is characterized in that the multi-modal image fusion module is used for aligning anatomical features of different images, eliminating respiratory motion artifacts and predicting the dose distribution of the different images; the dose distribution prediction module is used for rapidly predicting radiotherapy dose distribution based on anatomical features and historical data, the dynamic adaptive optimization module is used for monitoring anatomical changes in real time and dynamically adjusting a dose plan, and the prognosis model integration module is used for quantifying correlation between dose distribution and radioactive injury risks. High-precision alignment and respiratory motion artifact elimination of an anatomical structure are achieved through the multi-modal image fusion module, three-dimensional dose calculation is rapidly and accurately conducted through the dose distribution prediction module, organ displacement and deformation are effectively coped with through the dynamic self-adaptive optimization module through a real-time image monitoring and reinforcement learning algorithm, and the accuracy of the three-dimensional dose calculation is improved. And the prognosis model integration module constructs an individualized risk prediction model.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, and in particular to an AI-assisted intelligent radiotherapy dose optimization device. Background Technology

[0002] Radiotherapy is one of the "three pillars" of cancer treatment, with its core objective being to precisely eliminate tumors while maximizing the protection of surrounding healthy tissues. However, traditional radiotherapy suffers from drawbacks such as subjective target definition, time-consuming dose calculations, and the inability to dynamically adjust treatment plans, limiting the improvement of efficacy and toxicity control. In recent years, the integration of artificial intelligence technology, through breakthroughs in key technologies such as multimodal image fusion, deep learning dose prediction, dynamic adaptive optimization, and prognostic model integration, has driven the transformation of radiotherapy dose optimization from "experience-driven" to "data-driven," ushering in a new era of precision radiotherapy.

[0003] Radiation dose is a core element of radiotherapy, and its precise design and implementation directly determine the therapeutic effect on tumors and the protection of normal tissues. The entire radiation dose system is based on both physical dosimetry and biological effect models, achieving a balance between tumor killing and normal tissue protection through multi-dimensional parameter optimization. From the perspective of dose calculation, modern radiotherapy generally uses pencil beam convolution algorithms or Monte Carlo algorithms. The former calculates the dose distribution by integrating discretized ray paths, while the latter provides a more accurate dose deposition model by simulating the microscopic interaction between photons and electrons, and is particularly suitable for dose gradient calculation in complex geometries. The dose unit is Gray (Gy), where 1 Gy equals 1 joule / kilogram. In clinical practice, a fractionation pattern of 5 fractions per week is commonly used, and the total dose is adjusted according to the tumor type. For example, squamous cell carcinoma of the head and neck usually uses 70 Gy / 35 fractions, while prostate cancer may use 78 Gy / 39 fractions.

[0004] The first step in radiotherapy is the accurate identification of the tumor target area and organs at risk. Traditional methods rely on doctors manually delineating these areas, which is time-consuming and highly subjective. Dosage calculation is a core aspect of radiotherapy; traditional methods are time-consuming and have large errors in modeling dose distribution for complex anatomical structures. During radiotherapy, the patient's anatomy may change due to respiratory movements, tumor regression, and organ filling, causing the planned dose distribution to deviate from expectations. Traditional radiotherapy cannot dynamically adjust this; it can only expand the irradiation range to cover uncertainties, increasing the risk of damage to normal tissues. Therefore, this invention proposes an AI-assisted intelligent radiotherapy dose optimization device to address the problems existing in the prior art. Summary of the Invention

[0005] To address the aforementioned issues, the present invention aims to propose an AI-assisted intelligent radiotherapy dose optimization device. This device achieves high-precision alignment of anatomical structures and elimination of respiratory motion artifacts through a multimodal image fusion module, and improves the efficiency and consistency of target delineation by combining deep learning segmentation technology. The dose distribution prediction module performs rapid and accurate three-dimensional dose calculation, and the error compensation mechanism further ensures the feasibility of the treatment plan. The dynamic adaptive optimization module effectively addresses organ displacement and deformation through real-time image monitoring and reinforcement learning algorithms. The prognostic model integration module integrates multi-dimensional clinical data and genetic information to construct an individualized risk prediction model, providing a biomedical basis for dose optimization.

[0006] To achieve the objectives of this invention, the invention is implemented through the following technical solution: an AI-assisted intelligent optimization device for radiotherapy dose, comprising a multimodal image fusion module, a dose distribution prediction module, a dynamic adaptive optimization module, and a prognostic model integration module. The multimodal image fusion module is used to align the anatomical features of different images and eliminate respiratory motion artifacts. The dose distribution prediction module, based on anatomical features and historical data, quickly predicts the radiotherapy dose distribution. By inputting segmentation results and image data, it predicts the dose distribution of different treatment plans. The dynamic adaptive optimization module is used to monitor anatomical changes in real time and dynamically adjust the dose plan. The prognostic model integration module is used to quantify the correlation between dose distribution and the risk of radiation damage.

[0007] Further improvements are made in that the multimodal image fusion module includes a data preprocessing unit, a cross-modal registration unit, and an automatic segmentation unit.

[0008] A further improvement is that the data preprocessing unit is used to unify images of different resolutions to 2.5mm. 3 The data preprocessing unit receives CT images, magnetic resonance images, and metabolic activity images, and uses nonlocal mean filtering to eliminate noise.

[0009] Further improvements are made in that: the cross-modal registration unit is used to align the spatial coordinates of different modal images using rigid registration, and to adjust the local deformation field for organ deformation using elastic registration, thereby eliminating spatial resolution differences between different devices and aligning anatomical structures.

[0010] Further improvements are made in that: the automatic segmentation unit is used to segment the target area and organs at risk based on a deep learning network, extract image features, encode and output high-dimensional feature maps, eliminate respiratory motion artifacts, and generate dynamic tumor volumes.

[0011] A further improvement is that the cross-modal registration unit aligns the soft tissue contrast of the magnetic resonance image with the anatomical structure of the CT scan to generate a dynamic fused image.

[0012] Further improvements are made in that: the dose distribution prediction module includes an input feature extraction unit, a deep learning prediction unit, and an error compensation unit. The input feature extraction unit is used to input patient images and target area information and output a three-dimensional dose distribution. The deep learning prediction unit is based on a Transformer-GCN hybrid architecture, which captures long-range spatial dependencies through a self-attention mechanism, combines graph convolutional networks to model the association between dose distribution and anatomical structures, and outputs a predicted dose distribution as a supervision signal. The error compensation unit calculates the prediction error based on the predicted dose distribution output by the deep learning prediction unit.

[0013] Further improvements include: the input of the prognostic model integration module includes dose-volume parameters, patient age, tumor stage, and genomic data.

[0014] The beneficial effects of this invention are as follows: This invention achieves high-precision alignment of anatomical structures and elimination of respiratory motion artifacts through a multimodal image fusion module, and improves the efficiency and consistency of target area delineation by combining deep learning segmentation technology; the dose distribution prediction module performs fast and accurate three-dimensional dose calculation, and the error compensation mechanism further ensures the feasibility of the plan; the dynamic adaptive optimization module effectively addresses organ displacement and deformation through real-time image monitoring and reinforcement learning algorithms; the prognostic model integration module integrates multi-dimensional clinical data and genetic information to construct an individualized risk prediction model, providing a biomedical basis for dose regimen optimization. Attached Figure Description

[0015] Figure 1 This is a system architecture diagram of the present invention;

[0016] Figure 2 This is a diagram illustrating the architecture of the multimodal image fusion module of the present invention.

[0017] Figure 3 This is a diagram of the dose distribution prediction module architecture of the present invention. Detailed Implementation

[0018] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0019] In the dose optimization phase, intensity-modulated radiotherapy (IMRT) dynamically adjusts the radiation field intensity using a multi-leaf collimator, achieving uniform dose coverage within the tumor target area while reducing the radiation dose to surrounding organs at risk. Volumetric rotating intensity-modulated radiotherapy (VMAT) further optimizes gantry rotation speed and dose rate modulation, reducing treatment time to one-third of traditional IMRT. Image-guided radiotherapy (IGRT) uses cone-beam computed tomography (CBCT) or ultrasound to calibrate positioning errors in real time, ensuring that the actual irradiation position deviates from the planned position by less than 3 mm. Adaptive radiotherapy (ART) dynamically adjusts the dose distribution based on changes in images during treatment. For example, for prostate tumors, the system can automatically scale the radiation field boundaries or re-optimize dose weights due to anatomical displacement caused by changes in rectal filling. The dose verification system includes two dimensions: absolute dose measurement and relative dose verification. Absolute dose is measured in a solid water phantom using an ionization chamber, and the deviation from the calculated value by the planning system must be controlled within ±3%. Relative dose validation uses film dosimeters or array detectors to assess the spatial consistency of dose distribution. Gamma analysis quantifies the consistency between the planned and measured doses by comparing dose differences (typically set at 3% / 3mm). Emerging dose validation technologies, such as 3D dose printing and virtual dose validation systems, further enhance the intuitiveness and efficiency of validation. The introduction of biodosimetry has ushered in a personalized era for radiotherapy. The Normal Tissue Complication Probability (NTCP) model, using logistic regression or machine learning methods, integrates dose-volume parameters (such as maximum spinal cord dose and average parotid gland dose) with covariates such as patient age and underlying diseases to predict the risk of adverse reactions such as radiation pneumonitis and esophagitis. The Tumor Control Probability (TCP) model estimates the probability of local tumor control based on cell survival curve parameters (α / β ratio) and dose distribution. The combined therapeutic ratio provides a biological basis for optimal treatment planning. At the level of advanced radiotherapy technology, the Bragg peak characteristics of proton therapy enable dose sculpting, precisely delivering high-dose areas to deep tumor sites while significantly reducing the output dose. This has revolutionary implications for the treatment of pediatric tumors and tumors near vital organs. Heavy ion therapy (such as carbon ions), with its higher relative biological effect (RBE), has demonstrated superior killing effects in radiation-induced tumors. FLASH radiotherapy, through instantaneous irradiation at ultra-high dose rates (>40 Gy / s), has shown the potential to reduce damage to normal tissues in animal experiments, and its clinical translation is accelerating. The spatiotemporal heterogeneity of dose-effect relationships is currently a research hotspot. 4D dose calculation incorporates organ motion into the dose accumulation model, using Deformable Image Registration technology to track the spatiotemporal deformation of tumors and organs at risk, ensuring the accuracy of dose superposition between fractionated treatments. Immunoradiotherapy dose design is beginning to consider distance effects; low-dose radiotherapy combined with immune checkpoint inhibitors can activate systemic anti-tumor immunity. In this case, dose selection must balance the optimal window for local killing and systemic immune activation.A quality control system is implemented throughout the entire radiotherapy process. Daily morning checks verify the linear accelerator's output dose, energy stability, and mechanical precision using a morning check instrument. Weekly random checks of IMRT / VMAT plans are conducted using the ArcCheck 3D dose verification system. Monthly accelerator output dose calibration is performed, and a comprehensive QA program is conducted annually by the physicist team. The use of independent dose verification software (such as VeriSoft) further reduces the risk of human error. With the penetration of artificial intelligence, dose prediction models are undergoing a paradigm shift. Deep learning-based dose prediction systems, trained on millions of case data, can automatically generate initial dose plans that meet clinical goals, reducing traditional manual optimization time from hours to minutes. The application of reinforcement learning algorithms in adaptive radiotherapy allows for autonomous decision-making on the optimal dose adjustment strategy based on real-time image feedback, achieving true dynamic dose optimization.

[0020] Based on this, according to Figure 1 , Figure 2 , Figure 3 As shown, this embodiment provides an AI-assisted intelligent radiotherapy dose optimization device, including a multimodal image fusion module, a dose distribution prediction module, a dynamic adaptive optimization module, and a prognostic model integration module. The multimodal image fusion module is used to align the anatomical features of different images and eliminate respiratory motion artifacts. The multimodal image fusion module includes a data preprocessing unit, a cross-modal registration unit, and an automatic segmentation unit. The data preprocessing unit is used to unify images of different resolutions to 2.5mm. 3 The data preprocessing unit receives influencing data including CT images, MRI images, and metabolic activity images. The cross-modal registration unit uses rigid registration to align the spatial coordinates of images from different modalities and elastic registration to adjust the local deformation field for organ deformation, eliminating spatial resolution differences between different devices and aligning anatomical structures. In the cross-modal registration unit, the soft tissue contrast of MRI images is aligned with the anatomical structures of CT images to generate a dynamic fused image. The automatic segmentation unit uses a deep learning network to segment the target area and organs at risk, extract image features, encode and output high-dimensional feature maps, eliminate respiratory motion artifacts, and generate dynamic tumor volumes. This module integrates multi-source medical image data such as CT, MRI, and metabolic activity images through deep learning and cross-modal alignment technology to achieve precise anatomical structure localization and target area delineation. The input images are first resampled and standardized, then cross-modal feature extraction is performed using a 3D U-Net++ network, and a segmentation mask for the target area and organs at risk is generated using an encoder-decoder structure. Meanwhile, a dynamic attention mechanism is used to align the anatomical features of different images, and motion artifacts are eliminated through 4D-CT respiratory gating technology to generate dynamic fused images.

[0021] The dose distribution prediction module rapidly predicts radiotherapy dose distribution based on anatomical features and historical data. Inputting segmentation results and image data, it predicts dose distributions for different treatment plans. The module includes an input feature extraction unit, a deep learning prediction unit, and an error compensation unit. The input feature extraction unit takes patient images and target area information as input and outputs a 3D dose distribution. The deep learning prediction unit, based on a Transformer-GCN hybrid architecture, captures long-range spatial dependencies through a self-attention mechanism and models the relationship between dose distribution and anatomical structures using a graph convolutional network, outputting the predicted dose distribution as a supervisory signal. The error compensation unit calculates the prediction error based on the predicted dose distribution output by the deep learning prediction unit. Based on historical radiotherapy planning data and a deep learning model, the module rapidly predicts the 3D dose distribution. The input consists of segmented target area / OAR mask and image features. The Transformer-GCN hybrid architecture captures long-range spatial dependencies, generates a flux map, and then combines Monte Carlo simulation and pencil beam analysis to calculate the dose distribution. The model employs a two-stage training process: the first stage uses historical planning data as input to predict intermediate dose distributions; the second stage uses intermediate doses as input to optimize the final dose distribution. The prediction accuracy reaches ≤1.5Gy for D95 error of PTV and ≤3Gy for maximum dose error of OAR, and the deviation is corrected through an error compensation mechanism.

[0022] The dynamic adaptive optimization module is used to monitor anatomical changes in real time and dynamically adjust the dosage plan. It monitors anatomical changes during treatment and triggers dose re-optimization. The module acquires real-time images every 36 seconds and calculates the target displacement vector field using the Demons deformation registration algorithm. If the displacement is >3mm or the organ volume change is >5%, the optimization process is initiated. Optimization is based on reinforcement learning, using TCP / NTCP balance as the objective function, and adjusts subfield weights or beam angles.

[0023] The prognostic model integration module quantifies the association between dose distribution and the risk of radiation damage. Inputs to this module include dose-volume parameters, patient age, tumor stage, and genomic data. It predicts complication risk using a Cox proportional hazards model and generates a risk score. The model integrates a multi-center clinical database and employs transfer learning to adapt to the characteristics of different radiotherapy equipment.

[0024] The AI-assisted intelligent optimization device for radiotherapy dose first aligns CT, MRI, and metabolic activity images through a multimodal image fusion module to eliminate respiratory motion artifacts and segment the target area and organs at risk. Subsequently, the dose distribution prediction module predicts the three-dimensional dose distribution based on anatomical features and historical data, combined with a deep learning model. During treatment, the dynamic adaptive optimization module monitors anatomical changes in real time. When displacement or organ volume changes exceed a threshold, the optimization mechanism is triggered, and the radiotherapy plan is adjusted through reinforcement learning. Finally, the prognostic model integration module integrates dose parameters, patient clinical characteristics, and genetic data to quantify and predict the risk of radiation damage and generate a risk score, forming a closed-loop system covering the entire process from image processing and dose planning to real-time optimization and prognostic assessment.

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-assisted intelligent optimization device for radiotherapy dosage, characterized in that: The system includes a multimodal image fusion module, a dose distribution prediction module, a dynamic adaptive optimization module, and a prognostic model integration module. The multimodal image fusion module is used to align the anatomical features of different images and eliminate respiratory motion artifacts. The dose distribution prediction module quickly predicts the radiotherapy dose distribution based on anatomical features and historical data. By inputting segmentation results and image data, it predicts the dose distribution of different treatment plans. The dynamic adaptive optimization module is used to monitor anatomical changes in real time and dynamically adjust the dose plan. The prognostic model integration module is used to quantify the correlation between dose distribution and the risk of radiation damage.

2. The AI-assisted intelligent radiotherapy dose optimization device according to claim 1, characterized in that: The multimodal image fusion module includes a data preprocessing unit, a cross-modal registration unit, and an automatic segmentation unit.

3. The AI-assisted intelligent radiotherapy dose optimization device according to claim 2, characterized in that: The data preprocessing unit is used to unify images of different resolutions to 2.5mm. 3 The data preprocessing unit receives CT images, magnetic resonance images, and metabolic activity images, and uses nonlocal mean filtering to eliminate noise.

4. The AI-assisted intelligent radiotherapy dose optimization device according to claim 2, characterized in that: The cross-modal registration unit is used to align the spatial coordinates of images of different modalities using rigid registration and to adjust the local deformation field for organ deformation using elastic registration, thereby eliminating spatial resolution differences between different devices and aligning anatomical structures.

5. The AI-assisted intelligent radiotherapy dose optimization device according to claim 2, characterized in that: The automatic segmentation unit is used to segment the target area and organs at risk based on a deep learning network, extract image features, encode and output high-dimensional feature maps, eliminate respiratory motion artifacts, and generate dynamic tumor volume.

6. The AI-assisted intelligent radiotherapy dose optimization device according to claim 2, characterized in that: The cross-modal registration unit aligns the soft tissue contrast of the magnetic resonance imaging with the anatomical structure of the CT scan to generate a dynamic fused image.

7. The AI-assisted intelligent radiotherapy dose optimization device according to claim 1, characterized in that: The dose distribution prediction module includes an input feature extraction unit, a deep learning prediction unit, and an error compensation unit. The input feature extraction unit is used to input patient images and target area information and output a three-dimensional dose distribution. The deep learning prediction unit is based on a Transformer-GCN hybrid architecture, which captures long-range spatial dependencies through a self-attention mechanism and combines graph convolutional networks to model the relationship between dose distribution and anatomical structures, and outputs a predicted dose distribution as a supervision signal. The error compensation unit calculates the prediction error based on the predicted dose distribution output by the deep learning prediction unit.

8. The AI-assisted intelligent radiotherapy dose optimization device according to claim 1, characterized in that: The prognostic model integration module takes into account dose-volume parameters, patient age, tumor stage, and genomic data.

Images